Introduction
I have a rule I try never to break when a big model launches: I separate what I can actually use today from what the announcement wants me to believe. The gap between those two things is usually where the hype lives, and this week's launch from Tencent is a perfect example.
On November 4, 2026, Tencent released the full version of Hy4, its flagship Hunyuan model. It is the first model in the Hunyuan line that can see images and operate a computer — two capabilities that genuinely matter, and that put it in the same conversation as the computer-use agents from Anthropic and OpenAI. The headlines mostly call it a new open-source giant, a 770-billion-parameter model. Some of that is true. Some of it is not true yet. And the difference matters a lot if you are deciding whether to build on it.
What Actually Launched on November 4
Hy4 is not a brand-new model appearing out of nowhere. It is the full release of a model Tencent previewed in late August 2026. That preview was text-only: it could write code and handle office-style work, but it could not see an image at all. Ten weeks later, the full release adds the two things the preview lacked.
The scale is genuinely large. Hy4 is a mixture-of-experts model with 770 billion total parameters, up from 295 billion in Hy3 — roughly a 2.6x jump. Mixture-of-experts is worth understanding here, because the headline number misleads people. Only a fraction of those 770 billion parameters activates for any given token — 49 billion, according to the preview's published specs. That is how a model this size stays servable at all: the 770 billion is total capacity, not per-token compute, so do not compare it directly with a dense model's parameter count.
There is also a detail about the specs that deserves honesty. Tencent's launch announcement for the full release does not itself restate the parameter count or the context length. The 770 billion total, 49 billion active, and 1-million-token context window figures in circulation come from the August preview's specs and from API gateway listings like OpenRouter. They are almost certainly carried over, but strictly speaking they are preview figures, not restated launch specs.
What is confirmed about the launch itself is clear enough: Hy4 went live on November 4 through Tencent Cloud's TokenHub, through OpenRouter, and inside Tencent's own products — WorkBuddy and CodeBuddy, the Yuanbao assistant, ima, and the WeCom intelligent assistant Dayuan. You can use it today, as a hosted service, through an API.
The First Big Addition: Vision
The first new capability is visual understanding, and it is more than a checkbox feature. Hy4 can now process images at a fine-grained level, analyse complex charts, and reason about mathematical visuals — combining what it sees with deep reasoning and tool calls.
Tencent's own demonstration of this is striking, and I will flag it for what it is: a vendor illustration, not an independent benchmark. The demo uses a 2.8-gigapixel aerial photograph, and the model locates a two-metre roof component inside it. Impressive, and a hint of real capability for document analysis and visual inspection — but no published visual benchmark is attached to it, so treat it as a signpost, not proof.
Where vision actually matters for developers is less glamorous and more useful: triaging a bug report that arrives as a screenshot, extracting structured data from a chart in a PDF, or checking a design mockup against a live page. A model that can genuinely read a messy, real-world image — not just caption it vaguely — saves a whole round of back-and-forth. That is the practical test I would run on Hy4's vision first.
The Second Big Addition: Computer Use
The second new capability is the one that gets me most interested, and most cautious. Hy4 can now operate a computer through its graphical interface and command line, combining perception, deep reasoning, and tool invocation. In plain terms, it can look at a screen, click buttons, type, run commands, and carry a task across several applications.
Tencent did not build this for casual web browsing. The training focused on office tasks, financial analysis, and game development, including operating professional software like Blender and CAD tools, and adapting the model for game engines like Unreal Engine, Unity, and Godot, plus long tasks spanning several applications — the kind of multi-step work where most agents fall apart halfway through.
This puts Hy4 in a small, serious category. Anthropic's computer-use agents and OpenAI's operator-style agents are the obvious comparisons, and Tencent published a head-to-head number on the OSWorld 2.1 computer-use benchmark: Hy4 scored 78.9, against 81.8 for Claude Opus 5.5, 79.6 for Kimi K3.1 Pro, and 75.2 for GPT-6.1 Sol. Read that table the way it deserves to be read: it is Tencent's own run, not an independently reproduced result. Until someone outside Tencent runs OSWorld on this exact model, those comparisons are vendor-reported. They are encouraging, not settled.
My own view, after watching the computer-use race all year, is that the benchmark number matters less than reliability on long tasks. An agent that scores well on short, clean tasks but drifts or clicks the wrong button on step fourteen is not something you hand a production workflow. The honest test for Hy4 is a boring one: give it a real, multi-application task, watch every step, and count how often it needs rescuing. I would not let any computer-use agent — Hy4's or anyone else's — touch a live client site, a production database, or a payment flow unsupervised. The technology is exciting precisely because the failure modes are expensive.
The Open-Source Asterisk: The Weights Are Not Out Yet
Here is the part of the story that most headlines get wrong, and it is the part I most want you to take away.
The August preview shipped with its weights on day one, on Hugging Face, under the Apache 2.0 license. That is why the Hunyuan line earned its open reputation, alongside DeepSeek, Alibaba's Qwen, and Moonshot's Kimi. The full Hy4 release breaks that pattern. Tencent says it plans to release the model weights and a comprehensive system card, with its corporate release saying "in the coming days," but as of the launch, the weights are not downloadable, the licence for the full release has not been stated, and Tencent's own Hugging Face account still lists only the August preview. The "Model weights" button on OpenRouter's Hy4 page even links to a Hugging Face URL that is not publicly accessible.
Tencent is also unusually candid about the gap. Its launch post says Hy now ranks among the strongest open-source models "for the first time," while acknowledging a "large gap" between Hy4 and the top closed models. The fuller safety evaluations will only arrive with the system card, at the same time as the weights.
So, plainly: right now Hy4 is a service, not open weights. You can call it through an API, but you cannot download it, run it on your own servers, or fine-tune it — and if your plans depend on self-hosting, you are waiting. A lighter Hy4 Flash version is also coming soon, with no specs, date, or price yet.
Benchmarks and Numbers Worth a Second Look
Tencent published a table of 34 benchmarks pitting Hy4 against its preview, GLM-5.3, DeepSeek-V4.1-Flash, and Kimi K3.1 Pro, plus the closed models GPT-6.1 Sol and Claude Opus 5.5. Two results from Tencent's own evaluations are worth highlighting, with the usual vendor-reported caveat attached.
First, a finance evaluation built with 213 finance professionals, including 153 experts from financial institutions, across 579 real-world tasks. Tencent says Hy4 ranked first among open-weight models with a score of 3.04, ahead of DeepSeek-V4.1-Flash at 2.97 and Kimi K3.1 Pro and GLM-5.3 at 2.68 each. That is a genuinely interesting evaluation design — real tasks judged by real professionals, not just multiple-choice questions — and it fits with the model's office-and-finance training focus.
Second, hallucinations. Tencent reports a large reduction in its own search tests versus the preview — but its two English announcements disagree on the figures (15.76% to 5.45% in one, 12.95% to 7.2% in the other), so quote neither pair as gospel until an independent check exists.
Pricing, meanwhile, is unchanged from the preview: 6 yuan per million input tokens and 18 yuan per million output tokens, which works out to roughly $0.834 per million input tokens and $2.501 per million output tokens. There is a free quota that ends November 15, 2026, so if you want to try Hy4 without paying, this week is the window — and I would use it for evaluation, not for anything you cannot afford to rebuild later.
What This Means for WordPress and Elementor Builders
You might be wondering what a 770-billion-parameter Chinese flagship model has to do with your WordPress work. The most direct answer is about computer use:
First, computer-use agents are coming for the browser-based workflows we all live in — including WordPress dashboards, Elementor editing, and plugin settings. A model that can operate a GUI reliably will eventually be able to audit a site's settings, migrate content between pages, or test a checkout flow by actually clicking through it. That is powerful, and it is also a security story: if agents can use your site's admin interface, then your site's admin interface is now an agent interface too, whether you planned it or not. Strong authentication, least-privilege user roles, and staging environments are the guardrails for an agent era. I would not give a computer-use agent an administrator account on a live site today, full stop.
What I Would Do This Week
If you want to evaluate Hy4 sensibly, here is the short list I would actually follow.
Try it through the free quota before November 15, but keep the evaluation small and deliberate. Test vision on your own messy images — a real client screenshot, a dense chart, a scanned document — not on demo material. Test computer use on a disposable virtual machine or a staging environment, never on a production site, and watch every step. Compare its answers against a model you already trust on the same prompts, and keep score yourself rather than trusting a vendor table.
Do not plan a project around self-hosting Hy4 until the weights are actually downloadable and the licence is stated. If someone on your team says "we will just run Hy4 locally," ask them to show you the download link. Right now, there is not one for the full release.
Conclusion
Hy4 is a genuinely significant release with a genuinely misleading headline. The vision and computer-use capabilities are real, the scale is real, and the API is live today at a competitive price. But the open weights — the thing that would make it truly open — are still a promise, the headline specs are carried over from the preview, and the most impressive numbers are vendor-reported and, in the case of hallucinations, inconsistent across Tencent's own announcements.
That is not a reason to ignore Hy4. It is a reason to try it with your eyes open: use the free window, test it on your own work, and wait for the weights and the system card before you build anything that depends on them. The models that matter in the long run are not the ones with the loudest launch day. They are the ones that still look good after a month of real use, on real tasks, by people with no reason to flatter them. Hy4 has earned a place in that test. It has not passed it yet.
FAQ
Is Tencent's Hy4 really open source?
Not yet in practice. The August preview was released with open weights under Apache 2.0, but the full November 4 release is currently API-only. Tencent says it plans to release the weights and a system card, but no date has been given and the licence for the full release is unstated.
What is new in Hy4 compared to the preview?
Two major capabilities: visual understanding (fine-grained image processing, complex chart analysis, mathematical visual reasoning) and computer use (operating a computer through its graphical interface and command line). The preview was text-only.
How big is Hy4?
It is a mixture-of-experts model with 770 billion total parameters, 49 billion of which activate per token, according to the preview's specs — figures that Tencent's full-release announcement did not itself restate. Its context window is listed at 1 million tokens on API gateway listings.
How much does Hy4 cost to use?
Pricing is unchanged from the preview: about $0.834 per million input tokens and $2.501 per million output tokens (6 and 18 yuan respectively). A free quota runs until November 15, 2026.
Can Hy4 run on my own servers?
No, not yet. Until Tencent publishes the full release's weights, every request goes through Tencent's API via TokenHub, OpenRouter, or Tencent's own apps. A lighter Hy4 Flash version is also promised soon.
Should I let a computer-use agent touch my WordPress site?
Not on a live site, and not with an administrator account. Test computer-use agents only in staging or disposable environments, watch every step, and keep strong authentication and least-privilege roles in place. The technology is promising; unsupervised access to production is not worth the risk.
